Reproduce “TideCache: Efficient Hybrid LLM Serving with Transparent Cache Management and Layer-Wise Streaming” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinbin Luo, Xiaozhe Yao, Shang Yang, Zhijian Liu, Song Han, Qinghao Hu
TideCache serves hybrid models by dynamically reallocating GPU pages between global-KV and bounded-state caches while streaming historical prefixes from host memory layer by layer. It overlaps transfers with computation and keeps newly generated KV resident. On GPT-OSS-120B/GH200, it completes 2.04× as many agent turns as tuned SGLang with HiCache when prefixes exceed global-KV capacity, a serving result rather than less generation work.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.